| paper | d4-conf-aaai-2022-1385 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | no |
The paper’s appendix discusses gradient descent and gives a complexity observation, and another appendix lists algorithms, but the paper contains no numbered Theorem, Lemma, Corollary, or Proposition with a computational claim that can anchor the mirror ([full paper](https://arxiv.org/abs/2202.10153)). The proposed mixture over reward types changes LORI’s shared-reward inference target, and no result in the paper establishes that mixture problem.
fails bit a — no named computational result to mirror
LORI’s reward inference, gradient-based learning, and healthcare experiments are uncovered because none is stated as a qualifying numbered computational result; the mixture extension is outside the paper’s results.
The paper contains no qualifying named computational result: there is no numbered theorem, lemma, corollary, or proposition asserting a complexity classification. It proposes LORI, derives differentiability for its likelihood, and reports experiments, but makes no \(P\), NP-hardness, FPT, or similar claim. So there are no valid anchors here, and no per-anchor continuous problem or trichotomy prediction I can honestly provide.
A plausible *unanchored* extension would model many clinicians or decision-makers as a finite set \(T\) of types, where each type has a complete lexicographic reward model. Let \(\mu_t\) be the share of type \(t\), and let \(p_t(x\succ y)\) be that type’s comparison probability under the paper’s model. Given pooled pairwise preference counts \(n(x,y)\), infer \(\mu\) by maximizing the likelihood whose comparison probabilities are \(\sum_{t\in T}\mu_t p_t(x\succ y)\). The solution is a distribution over decision-maker types.
That population setting has some support in the paper: it explicitly allows observed preferences to come from a population of decision-makers. But LORI fits a shared set of reward functions with stochastic comparisons; inferring a mixture of heterogeneous reward types changes the inference target. That is the weakest point of the positive case, and this extension cannot substitute for the missing named computational anchor.
The negative case starts at the gate: in the supplied text, there is no named result classifying the complexity of a computational problem. LORI is a preference inference method, with a likelihood objective, a gradient descent procedure for a parameterized family, and experiments. There is no theorem or algorithmic complexity claim here to carry into a continuous-population version.
The proponent’s best extension is a mixture over clinician or expert types. For a finite type set \(T\), assign each type mass \(\mu_t\), with pairwise probability \(q_{xy}=\sum_{t\in T}\mu_t p_t(x\succ y)\), and infer the masses from pooled comparisons. That is a plausible population model, but it changes the target: LORI estimates shared lexicographic reward functions, while the mixture estimates a distribution of different reward models. The paper allows a population to generate observed comparisons, but its likelihood still uses one shared reward model; its stochastic comparison probabilities can already represent population-level variation without inferring latent types.
Keeping that shared-model target does not supply the missing mirror either. The observations already enter through pairwise counts \(n(x,y)\); replacing counts by proportions describes the data continuously, but does not turn the society into a distribution over voter types in the programme’s sense. In the transplantation example, treating patients as that society would change the task again: patients and organs are alternatives in individual allocation decisions, and modeling their masses would introduce allocation constraints the paper does not study.
So the strongest case against is a scope argument, not proof that no adjacent project could be worthwhile. A continuous mixture model of decision-makers may be worth studying on its own; it is not a continuous mirror of a computational result in this paper. The negative case is weak against a broader programme willing to take on neighboring inference problems, but under the stated standard for mirroring this paper, there is no surviving anchor.
The adversarial triple: the proponent anchors on up to three named results; the opponent sees that case and must defeat every anchor; the judge decides which case convinced it. These are the pipeline’s own outputs, generated by tools/triple_run.py — no human edited them. The paper’s own text is not reproduced here beyond the quoted statement above.